Linear Classifier
A linear classifier is a lightweight model component that takes a model's internal representation as input and outputs class probabilities through a linear transformation followed by an activation function. It is commonly used in probe analysis and inference-time scheduling.
Timeline (last 90 days)
AmbiSteer introduces a linear classifier to predict whether an input is ambiguous and conditions the steering direction on this.
The linear classifier is a lightweight probe that takes the model's internal hidden states as input and outputs an ambiguity probability through one linear transformation plus an activation function, with very few parameters and low training data requirements.
The classifier's accuracy directly determines the quality of the scheduling: a missed detection leaves ambiguous inputs unsteered, while a false positive applies unnecessary steering to unambiguous inputs.
The evaluation used a linear classifier with Softmax output, and ensured that the signal-level data splits were completely consistent.
The current evaluation of FA-PI is limited to linear classifiers, and whether more complex classifiers can further amplify the advantages of FA-PI remains to be explored.
All Facts (5)
AmbiSteer introduces a linear classifier to predict whether an input is ambiguous and conditions the steering direction on this.
50%UnverifiedThe linear classifier is a lightweight probe that takes the model's internal hidden states as input and outputs an ambiguity probability through one linear transformation plus an activation function, with very few parameters and low training data requirements.
50%UnverifiedThe classifier's accuracy directly determines the quality of the scheduling: a missed detection leaves ambiguous inputs unsteered, while a false positive applies unnecessary steering to unambiguous inputs.
50%UnverifiedThe evaluation used a linear classifier with Softmax output, and ensured that the signal-level data splits were completely consistent.
50%UnverifiedThe current evaluation of FA-PI is limited to linear classifiers, and whether more complex classifiers can further amplify the advantages of FA-PI remains to be explored.
50%